Image defogging enhancement and intelligent perception collaborative optimization method based on flow matching

Abstract

The application provides a flow matching-based image defogging enhancement and intelligent perception collaborative optimization method, including the steps of: acquiring a foggy input image, constructing a common differential equation defining image transformation; inputting the input image into a fog perception vector field, the fog perception vector field including: an atmospheric scattering purifier and a defogging perception color lookup table, the atmospheric scattering purifier being used to extract multi-scale features from the foggy input image to generate a clearer output, and the defogging perception color lookup table being used to adaptively adjust the image color through a nonlinear color transformation mechanism; solving the common differential equation through an RK4 solver, and iteratively updating the foggy input image at each time step to ensure that the input image can be smoothly and stably transformed from the initial foggy state to a clear image. The application solves the technical problem that the existing method in the prior art is forced to reduce the model capacity due to the calculation efficiency constraint in the ultra-high-definition image defogging, resulting in incomplete fog concentration estimation and inability to balance the perception quality and calculation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-definition image defogging, in particular to an image defogging enhancement and intelligent perception collaborative optimization method based on flow matching. BACKGROUND

[0002] With the popularization of 4K imaging devices, existing defogging methods face a double dilemma in resolution adaptability: traditional algorithms based on physical models (such as atmospheric scattering models) are prone to color distortion and artifacts under complex atmospheric conditions, making it difficult to meet the stringent requirements of ultra-high-definition images on texture details; while data-driven deep learning methods can improve defogging quality, but their parameter quantity and computational complexity increase exponentially with resolution, resulting in GPU memory usage exceeding 12GB and frame rate less than 30fps when processing a single 4K image, which severely restricts the actual deployment value. More importantly, existing methods often have to reduce model capacity due to computational efficiency constraints in ultra-high-definition image defogging, resulting in incomplete fog concentration estimation and thus reducing the accuracy of downstream visual tasks such as target detection and semantic segmentation. Therefore, constructing an ultra-high-definition defogging framework that balances perception quality and computational efficiency has become an urgent challenge in the field of computer vision.

[0003] Currently, existing ultra-high-definition defogging methods, such as 4Kdehazing, 4K-Haze, and LapDehazeNet, mainly rely on traditional convolutional neural networks, which have a large number of model parameters. For example, the trainable parameters of 4KDehazing reach 34.55M, which to some extent increases the computational complexity and inference time. Although existing general models for UHD image restoration, such as UHDformer and UHDDIP, have smaller parameter quantities (0.34M and 0.81M, respectively), they rely on additional prior information such as normal maps, and their defogging effect is not as good as models specifically designed for the defogging task. SUMMARY

[0004] The present application aims to overcome the above technical deficiencies and provide an image defogging enhancement and intelligent perception collaborative optimization method based on flow matching, which solves the technical problem that existing methods in the prior art often have to reduce model capacity due to computational efficiency constraints in ultra-high-definition image defogging, resulting in incomplete fog concentration estimation and thus reducing the accuracy of downstream visual tasks such as target detection and semantic segmentation, and failing to balance perception quality and computational efficiency.

[0005] To achieve the above technical purpose, the technical solution of the present application provides an image defogging enhancement and intelligent perception collaborative optimization method based on flow matching, comprising the steps of:

[0006] obtaining a foggy input image, constructing a differential equation defining image transformation, the differential equation being defined by a fog perception vector field;

[0007] inputting the input image into a fog-aware vector field, the fog-aware vector field comprising: an atmospheric scattering purifier for extracting multi-scale features from the foggy input image to generate a clearer output, and a de-fog-aware color lookup table for adaptively adjusting image colors through a nonlinear color transformation mechanism;

[0008] solving the ordinary differential equation through an RK4 solver, iteratively updating the foggy input image at each time step to ensure that the input image can smoothly and stably transition from the initial foggy state to a clear image.

[0009] Compared with the prior art, the beneficial effects of the present application include:

[0010] Flow Matching provides an efficient solution by directly learning the mapping relationship between data distributions, avoiding the complex iterative process in traditional generative models, and significantly reducing computational complexity and memory usage. When processing 4K images, the flow matching method of the present application reduces the parameter amount to 7.9M and the memory requirement to 3.2GB, significantly improving the computing efficiency. In addition, flow matching can efficiently model the complex distribution of high-resolution images without relying on additional prior information, thereby achieving better detail preservation and color restoration capability in the de-fogging task. Compared with UHDformer and UHDDIP, flow matching improves the PSNR index by about 2dB, achieving a better balance between performance and efficiency. Therefore, the present application attempts to introduce this technology to solve the challenge of super high definition de-fogging.

[0011] The fog removal based on flow matching technology is different from the strategies of progressive fog removal and diffusion fog removal. The progressive image dehazing method restores the haze-free image by optimizing in stages. This method usually uses multi-scale feature extraction, iterative optimization or hierarchical processing strategy to gradually refine the dehazing result, effectively alleviating the problem of detail loss or over-dehazing in a single dehazing process. However, the progressive image dehazing method usually relies on discrete iteration to gradually optimize the dehazing result through a preset discrete step. This fixed path may cause the algorithm to fall into a local optimum, especially in complex scenes or variable fog conditions, making it difficult to find a global optimal solution. At the same time, the progressive optimization based on physical priori lacks the ability to adaptively model complex fog distribution, and is prone to color shift problems in non-uniform fog, multi-scale fog cluster and other scenes. The diffusion model dehazing method can optimize the dehazing result in a more flexible and global way by simulating the natural diffusion process of data distribution. However, the reverse diffusion process requires hundreds of steps of iterative calculation, resulting in a high inference time of dozens of seconds for a single 4K image. In addition, the diffusion model is sensitive to noise, which can easily cause content and color distortion in low-texture areas. Although some studies have proposed accelerated sampling strategies, existing solutions still have significant bottlenecks in the trade-off between model lightweight and fidelity.

[0012] This application proposes an ultra-high-resolution image dehazing method based on flow matching and a haze-aware color lookup table (Haze-LUT), named 4KDehazeFlow. Unlike traditional dehazing methods, flow matching models the dehazing process as solving an ordinary differential equation (ODE) of a continuous vector field. By directly learning the mapping relationship from the hazy image to the clear image, it avoids the local optima problem caused by discretization iteration. Furthermore, this application designs a learnable 3D lookup table module, utilizing a data-driven adaptive nonlinear color transformation mechanism to achieve refined dehazing effects while ensuring efficient computation. Specifically, this application constructs a general flow matching dehazing framework, with core components including an initial distribution, a target distribution, a vector field, and an ODE solver. First, the hazy image is defined as the initial distribution, and the clear image is defined as the target distribution. The purpose is to gradually map the initial distribution to the target distribution using flow matching technology. Second, during the flow matching process, the vector field defines the transformation path for image dehazing. To further optimize the dehazing effect, this application introduces an improved CNN network into the vector field to extract multi-scale features of the image. Combined with a color-aware dehazing lookup table (Haze-LUT) module, it adaptively adjusts the image's color and brightness through a nonlinear color transformation mechanism, thereby improving the accuracy and visual quality of the dehazing effect. Finally, this application employs a fourth-order Runge-Kutta (RK4) differential equation solver, using an accurate step-by-step iterative method to stably solve the dehazing flow field. The RK4 method effectively balances computational accuracy and stability by calculating the weighted average of the four stages at each time step, ensuring the precision and reliability of the dehazing process. In the inference stage, based on the learned vector field and the dehazing-aware color lookup table (Haze-LUT), the input data is mapped to the target distribution by inversely solving ordinary differential equations, thereby achieving accurate image dehazing.

[0013] According to some embodiments of the present invention, the defogging process is controlled by the ordinary differential equation:

[0014] ;

[0015] in Indicates the time step. It is time The image, It is by Parameterized fog sensing vector field, It is the initial time step. For the input image, by solving the ordinary differential equation, the input image will gradually change from a foggy state to a clear state along a probability path.

[0016] According to some embodiments of the present application, the processing procedure of the haze-aware vector field on the input image is represented as:

[0017]

[0018] wherein, represents the output of the haze-aware vector field, is the current input image, is the output image obtained by the atmospheric scattering purifier, is the result of color adjustment by the haze-aware color lookup table, is a scaling factor to control the intensity of color transformation.

[0019] According to some embodiments of the present application, an atmospheric scattering model is designed in the atmospheric scattering purifier, represented as:

[0020]

[0021] wherein, is the clear image, is the input hazy image, is a learnable scaling parameter, is a learnable bias term with an initial value of 1.

[0022] According to some embodiments of the present application, a convolutional neural network (CNN) is used to directly estimate from the input. The dehazed image can be obtained by:

[0023]

[0024] wherein, is the current input image, is the output of the enhanced convolutional neural network. The convolutional neural network (CNN) is composed of three down-sampling encoder modules, one spatial attention module and three up-sampling decoder modules:

[0025]

[0026] wherein, is the current input image, represents the down-sampling encoder module, refers to the attention mechanism, represents the up-sampling decoder module.

[0027] According to some embodiments of the present application, a three-dimensional lookup table represents a three-dimensional lattice, composed of elements, denoted as wherein, is the number of quantization levels of each color channel.​​​​

[0028] each element corresponding to the RGB color index and its transformed output RGB color The accuracy of such a transformation is determined by The transformation of the haze-aware color lookup table is achieved through two key steps: lookup and trilinear interpolation.

[0029] According to some embodiments of the present application, for an input RGB color , the coordinates in the lookup table are calculated as follows:

[0030] ;

[0031] where is the maximum color value; after determining the position, the output RGB color is obtained by trilinear interpolation from the 8 adjacent lattice vertices; finally, the haze-aware color lookup table process is represented as:

[0032] ;

[0033] where is the input image, represents the trilinear interpolation function, is the haze-aware color lookup table containing the sampled output RGB values.

[0034] In a second aspect, the present application provides a super high-definition image dehazing system based on flow matching, comprising:

[0035] a data acquisition module for acquiring a hazy input image;

[0036] an equation construction module for constructing an ordinary differential equation defining image transformation, the ordinary differential equation being defined by a haze-aware vector field;

[0037] a haze-aware vector field, comprising: an atmospheric scattering purifier for extracting multi-scale features from the hazy input image to generate a clearer output, and a haze-aware color lookup table for adaptively adjusting image colors through a nonlinear color transformation mechanism;

[0038] an image dehazing module for solving the ordinary differential equation through an RK4 solver, iteratively updating the hazy input image at each time step to ensure that the input image can smoothly and stably transform from the initial hazy state to a clear image.

[0039] In a third aspect, the present application provides a computer readable storage medium, the computer readable storage medium storing computer executable instructions for causing a computer to execute the flow matching based image defogging enhancement and intelligent perception collaborative optimization method according to any one of the first aspect.

[0040] Additional aspects and advantages of the present application will be made apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0041] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0042] Fig. 1 A flow chart of the flow matching based image defogging enhancement and intelligent perception collaborative optimization method according to an embodiment of the present application is provided;

[0043] Fig. 2 A flow chart of the flow matching based image defogging enhancement and intelligent perception collaborative optimization method according to another embodiment of the present application is provided;

[0044] Fig. 3 A vector field schematic diagram of the flow matching based image defogging enhancement and intelligent perception collaborative optimization method according to another embodiment of the present application is provided. DETAILED DESCRIPTION

[0045] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0046] It should be noted that although the functional modules are divided in the system schematic diagram, and the logical order is shown in the flow chart, in some cases, the steps shown or described can be executed in a manner different from the module division in the system or the order in the flow chart. The terms "first", "second", etc. in the specification and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0047] Reference Figs. 1 to 3 , Fig. 1 A flow chart of the flow matching based image defogging enhancement and intelligent perception collaborative optimization method according to an embodiment of the present application is provided; Fig. 2 A flow chart of the flow matching based image defogging enhancement and intelligent perception collaborative optimization method according to another embodiment of the present application is provided; Fig. 3A vector field diagram of the flow matching based image defogging enhancement and intelligent perception collaborative optimization method is provided for another embodiment of the present application. The flow matching based image defogging enhancement and intelligent perception collaborative optimization method includes but is not limited to the following steps:

[0048] Step S110, an input image with fog is obtained, and an ordinary differential equation defining image transformation is constructed, the ordinary differential equation being defined by a fog perception vector field;

[0049] Step S110, the input image is input to the fog perception vector field, the fog perception vector field including: an atmospheric scattering purifier and a defogging perception color lookup table, the atmospheric scattering purifier being used to extract multi-scale features from the input image with fog to generate a clearer output, and the defogging perception color lookup table being used to adaptively adjust image colors through a nonlinear color transformation mechanism.

[0050] Step S110, the ordinary differential equation is solved by an RK4 solver, and the input image with fog is iteratively updated at each time step, so that the input image is smoothly and stably transformed from an initial foggy state to a clear image.

[0051] In an embodiment, the flow matching based image defogging enhancement and intelligent perception collaborative optimization method includes the steps of: obtaining an input image with fog, constructing an ordinary differential equation defining image transformation, the ordinary differential equation being defined by a fog perception vector field; inputting the input image to the fog perception vector field, the fog perception vector field including: an atmospheric scattering purifier and a defogging perception color lookup table, the atmospheric scattering purifier being used to extract multi-scale features from the input image with fog to generate a clearer output, and the defogging perception color lookup table being used to adaptively adjust image colors through a nonlinear color transformation mechanism; solving the ordinary differential equation by an RK4 solver, and iteratively updating the input image with fog at each time step, so as to ensure that the input image is smoothly and stably transformed from an initial foggy state to a clear image.

[0052] The atmospheric scattering purifier can extract multi-scale features from the foggy input image to generate a clearer output. The extraction of multi-scale features can capture information at different resolutions and different levels of detail, helping to remove the effects of fog in all directions. For example, for fog in the long shot and close-up, the corresponding features at different scales are different, and through the purifier, these differences can be accurately analyzed and processed to better restore the original clear appearance of the image, so that the dehazed image not only retains the overall outline, but also reveals rich details. The dehazing perceptual color lookup table uses a nonlinear color transformation mechanism to adaptively adjust the image color. In the dehazing process, fog often causes image color distortion, reduced contrast, and other problems, while the lookup table can adaptively change the color according to the actual dehazing situation of the image, avoiding the color incoordination and unnatural phenomena that may occur in traditional fixed color adjustment mode, making the dehazed image color more real and bright, consistent with human visual perception, and improving the visual quality of the image.

[0053] The RK4 solver is used to solve the ordinary differential equation, and the foggy input image is iteratively updated at each time step, which can ensure that the input image smoothly and stably changes from the initial foggy state to a clear image. It avoids image mutations and local over-processing that may occur during the dehazing process, making the dehazing effect more uniform and natural, and the entire image transition process gradual, resulting in a reliable clear image quality. In practical application scenarios, whether it is a static ultra-high definition image or a continuous video frame image (processed by frames), it can have stable and high-quality dehazing performance.

[0054] This application proposes an ultra-high-resolution image dehazing method based on flow matching and a haze-aware color lookup table (Haze-LUT), named 4KDehazeFlow. Unlike traditional dehazing methods, flow matching models the dehazing process as solving an ordinary differential equation (ODE) of a continuous vector field. By directly learning the mapping relationship from the hazy image to the clear image, it avoids the local optima problem caused by discretization iteration. Furthermore, this application designs a learnable 3D lookup table module, utilizing a data-driven adaptive nonlinear color transformation mechanism to achieve refined dehazing effects while ensuring efficient computation. Specifically, this application constructs a general flow matching dehazing framework, with core components including an initial distribution, a target distribution, a vector field, and an ODE solver. First, the hazy image is defined as the initial distribution, and the clear image is defined as the target distribution. The purpose is to gradually map the initial distribution to the target distribution using flow matching technology. Second, during the flow matching process, the vector field defines the transformation path for image dehazing. To further optimize the dehazing effect, this application introduces an improved CNN network into the vector field to extract multi-scale features of the image. Combined with a color-aware dehazing lookup table (Haze-LUT) module, it adaptively adjusts the image's color and brightness through a nonlinear color transformation mechanism, thereby improving the accuracy and visual quality of the dehazing effect. Finally, this application employs a fourth-order Runge-Kutta (RK4) differential equation solver, using an accurate step-by-step iterative method to stably solve the dehazing flow field. The RK4 method effectively balances computational accuracy and stability by calculating the weighted average of the four stages at each time step, ensuring the precision and reliability of the dehazing process. In the inference stage, based on the learned vector field and the dehazing-aware color lookup table (Haze-LUT), the input data is mapped to the target distribution by inversely solving ordinary differential equations, thereby achieving accurate image dehazing.

[0055] Among them, vector field It consists of an atmospheric scattering purifier and a dehazing sensing color lookup table, which are used to extract multi-scale features and perform adaptive color correction. The dehazing process is controlled by ordinary differential equations.

[0056] ;

[0057] in Indicates the time step. It is time The image, It is by Parameterized fog sensing vector field, It is the initial time step. For the input image, by solving the ordinary differential equation, the input image will gradually change from a foggy state to a clear state along a probability path.

[0058] In the flow matching process, the vector field defines the transformation path of image defogging, modeling the dynamic change of image distribution from foggy to clear over time. To achieve this goal, an atmospheric scattering purifier is developed inspired by AODNet, which directly generates clear images through a lightweight convolutional neural network (CNN). This method avoids separate transmission and atmospheric light estimation, can capture multi-scale features, and improve the defogging effect. However, since flow matching can perform defogging functions in multiple steps, it may cause image color loss. To further improve color accuracy, we propose a defogging-aware color lookup table, which is a three-dimensional lookup table (LUT) that adaptively adjusts image colors through a nonlinear color transformation mechanism. The processing process of the fog-aware vector field on the input image is represented as:

[0059] ;

[0060] where, represents the output of the fog-aware vector field, is the current input image, is the output image obtained by the atmospheric scattering purifier, is the result of color adjustment through the defogging-aware color lookup table, is the scaling factor that controls the color transformation intensity.

[0061] Atmospheric scattering purifier:

[0062] The atmospheric scattering purifier in 4KDehazeFlow plays a key role in extracting multi-scale features from foggy input images to generate clearer outputs. This convolutional neural network directly generates clear images without explicit estimation of the transmission map or atmospheric light, providing a lightweight and efficient defogging solution. The atmospheric scattering purifier is designed with an atmospheric scattering model, represented as:

[0063] ;

[0064] where, is the clear image, is the input foggy image, is a learnable scaling parameter, is a learnable bias term with an initial value of 1.

[0065] Through a convolutional neural network (CNN) for direct estimation of from the input, the defogged image can be obtained as follows:

[0066] ;

[0067] where, is the current input image, is the output of the enhanced convolutional neural network, which consists of three down-sampling encoder modules, one spatial attention module, and three up-sampling decoder modules:

[0068] ;

[0069] where, is the current input image, denotes the down-sampling encoder module, refers to the attention mechanism, indicates the up-sampling decoder module. The encoder extracts hierarchical features while progressively reducing spatial dimensions through max-pooling, convolution, adaptive normalization, and Gaussian error linear unit (GELU) activation functions. The spatial attention module highlights important regions, while the decoder optimizes features through bilinear up-sampling and skip connections to restore fine details. The network uses convolutional kernels of size 1x1 or 3x3.

[0070] Dehaze-aware color lookup table:

[0071] While the atmospheric scattering purifier can effectively remove haze, it often causes color loss or distortion in the process. To alleviate these problems, we propose a dehaze-aware color lookup table, which is a three-dimensional lookup table (LUT) that adaptively adjusts image colors through a nonlinear color transformation mechanism. A three-dimensional lookup table represents a three-dimensional lattice composed of elements, denoted as , where is the number of quantization levels for each color channel. Each element corresponds to the RGB color index and its transformed output RGB color . The transformation accuracy is determined by , which is usually set to 33 in practical applications. The transformation of the dehaze-aware color lookup table is achieved through two key steps: lookup and trilinear interpolation. For the input RGB color , the coordinates in the lookup table are calculated as follows:

[0072] ;

[0073] where, is the maximum color value. After determining the position, the output RGB color is obtained through trilinear interpolation from the eight adjacent lattice vertices. Finally, the dehaze-aware color lookup table process is represented as:

[0074] ;

[0075] wherein, is the input image, denotes a trilinear interpolation function, is a defogging-aware color lookup table containing the sampled output RGB values.

[0076] Improved effects:

[0077] The present application conducts comprehensive experiments on ultra-high definition (UHD) and non-ultra-high definition (non-UHD) datasets to evaluate the performance of the proposed method. The present application compares the method with seven state-of-the-art image defogging techniques, including Dehamer, C2PNet, MB-TaylorFormer, PTTD, 4KDehazing, UHDformer and UHDDIP.

[0078] The present application uses four widely recognized indicators to evaluate performance: PSNR, SSIM, LPIPS and NIQE. PSNR and SSIM focus on evaluating the fidelity and structural integrity of the reconstructed image, while LPIPS evaluates perceptual quality by measuring the similarity of deep features, providing a more human visual quality evaluation. NIQE is a no-reference image quality evaluation indicator that measures the distortion of an image by statistically analyzing the natural scene statistical characteristics of the image, and can evaluate the naturalness and authenticity of the defogging result without a reference image.

[0079] wherein, PSNR (Peak Signal-to-Noise Ratio) is a commonly used indicator for measuring image quality, mainly used to evaluate the difference between the processed image and the original image, the larger the value, the better the image quality, and the calculation formula is:

[0080] ;

[0081] ;

[0082] wherein, MSE represents the mean square error, which is used to measure the difference between two images at the pixel level, and respectively represent the pixel values of the generated fusion image and the corresponding ground truth image at position , m and n respectively represent the width and height of the image, and MAX represents the maximum value of the image pixel.

[0083] SSIM (Structural Similarity Index Measure) is an index for evaluating the quality of two images, mainly by comparing the brightness, contrast and structural information of the images to measure the similarity of the images. The closer the value is to 1, the better the image quality.

[0084] The calculation formula of SSIM is:

[0085] ;

[0086] ;

[0087] ;

[0088] wherein, is the average value of the fusion image output by the model , is the average value of the gold standard corresponding to , is the covariance representing the images and , is the variance of , is the variance of ; is the dynamic range of pixel value, and represent the preset hyperparameters, here =0.01, =0.03, and represent the smoothing parameters.

[0089] LPIPS (Learned Perceptual Image Patch Similarity) is a metric method for evaluating the perceptual similarity between two images. It is based on the features extracted by a deep convolutional neural network, rather than simply the pixel difference. The smaller the value, the better the image quality.

[0090] The calculation formula of LPIPS is:

[0091] ;

[0092] wherein, represents the fusion image output by the model, represents the gold standard image corresponding to and represent the first 4-layer features extracted by a pre-trained VGG deep neural network. layer feature map. is the image In the first layer feature, is the image In the first layer feature. represents the Euclidean distance (L2 norm) between feature maps. This is a difference measure between and , used to measure the similarity of two images in the layer feature. represents the weighted coefficient of each layer feature map. Different network layers have different effects on the perception of similarity, and the weight is used to reflect this, usually learned through training. represents the sum of the results of all layers, taking into account the feature differences at different levels.

[0093] NIQE (Natural Image Quality Evaluator) is a no-reference image quality evaluation index, used to measure the naturalness and distortion of an image. It is based on the statistical characteristics of natural images to construct a quality model, rather than relying on reference images for comparison. The smaller the value, the higher the image quality.

[0094] The calculation formula of NIQE is:

[0095] ;

[0096] where, is the feature mean vector of natural images (reference model), is the covariance matrix of natural images (reference model), is the feature mean vector of the image to be evaluated, is the covariance matrix of the image to be evaluated, is the transpose operation of the matrix, is the inverse operation of the matrix, is the average of the covariance matrix.

[0097] Table 1 Quantitative comparison on ultra-high definition dataset

[0098]

[0099] As can be seen from Table 1, the method 4KDehazeFlow of the present application achieves the best performance on the 4KID dataset, reaching the highest PSNR and the lowest LPIPS. Among them, the PSNR is improved by 0.11 dB compared with the suboptimal method, and the LPIPS is reduced by 0.012, indicating that the dehazed image has improved in both reconstruction quality and perceptual quality. On the I-HAZE dataset, the method of the present application achieves the optimal results in terms of PSNR, SSIM and LPIPS, further verifying its robustness and superiority on different datasets. In the O-HAZE dataset, 4KDehazeFlow performs best in terms of SSIM and LPIPS, with SSIM being 0.1755 higher than the suboptimal method and LPIPS being 0.0473 lower, indicating that it has obvious advantages in maintaining image structural integrity and perceptual quality. Overall, on the three ultra-high-definition datasets, the method 4KDehazeFlow of the present application achieves the optimal results in terms of average PSNR, SSIM and LPIPS, fully proving that its dehazing effect in ultra-high-definition scenes is superior to existing advanced methods.

[0100] Table 2 Quantitative comparison on SOTS dataset

[0101]

[0102] In order to verify the dehazing effect of the method of the present application in non-ultra-high-definition scenes, the present application further conducts experiments on the low-resolution dataset SOTS. The experimental results are shown in Table 2, and 4KDehazeFlow achieves the best performance in terms of NIQE and LPIPS, indicating that the dehazed image is superior to other methods in terms of naturalness and perceptual quality. In addition, in terms of SSIM, the method of the present application also performs well, only slightly lower than the MB-TaylorFormer method by 0.0101, still maintaining a high structural fidelity. These results further prove the robustness and superiority of 4KDehazeFlow in different resolution scenarios, not only suitable for ultra-high-definition datasets, but also competitive on low-resolution datasets.

[0103] Table 3 Comparison of computational efficiency of ultra-high-definition methods

[0104]

[0105] The complexity of the model is also evaluated in the present application, and Table 3 shows the comparison of the computational overhead of different ultra-high-definition (UHD) dehazing methods. Compared with UHDformer and UHDDIP, 4KDehazeFlow significantly reduces the number of multiply-accumulate operations (MACs) while maintaining competitive efficiency, effectively reducing the computational cost. Notably, the method of the present application can flexibly adjust the number of steps in flow matching, thereby balancing the inference time and performance requirements. Therefore, 4KDehazeFlow achieves a better trade-off between computational efficiency and dehazing quality.

[0106] In an embodiment, the flow matching based ultra-high-definition image dehazing system comprises: a data acquisition module for acquiring a foggy input image; a model construction module for constructing an ordinary differential equation defining image transformation, the ordinary differential equation being defined by a fog perception vector field; the fog perception vector field comprising: an atmospheric scattering purifier for extracting multi-scale features from the foggy input image to generate a clearer output, and a dehazing perception color lookup table for adaptively adjusting image color through a nonlinear color transformation mechanism; and an image dehazing module for solving the ordinary differential equation through an RK4 solver, iteratively updating the foggy input image at each time step to ensure that the input image can smoothly and stably transform from the initial foggy state to a clear image.

[0107] The memory, as a kind of non-transient computer readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs.In addition, the memory can include high-speed random access memory, and can also include non-transient memory, such as at least one magnetic disk storage device, flash memory device, or other non-transient solid-state storage device.In some embodiments, the memory can optionally include memory that is remotely arranged relative to the processor, and these remote memories can be connected to the processor through a network.The examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0108] The device embodiments described above are only schematic, wherein the units described as separate components can or can not be physically separated, and can be located in one place or distributed on multiple network units.A part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0109] In addition, one embodiment of the present application also provides a computer readable storage medium storing computer executable instructions, which are executed by a processor or controller, for example, by a processor in the above-mentioned terminal embodiment, so that the above-mentioned processor executes the flow matching based image dehazing enhancement and intelligent perception collaborative optimization method in the above-mentioned embodiment.

[0110] Those skilled in the art will appreciate that all or certain steps of the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Certain physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software can be distributed on computer readable media, which can include computer storage media (or non-transitory media) and communication media (or transitory media). As is well known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer. Further, it is common in the art for a person skilled in the art to appreciate that communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. Accordingly, the disclosure is considered to include any modification or derivative of the above disclosed methods that otherwise falls within the scope of the present claims.

[0111] The above description is only specific implementation of the present application, but the present application is not limited to the above implementation, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present application, and these equivalent modifications or replacements are included in the scope defined by the claims of the present application.

[0112] The above description of the specific implementation of the present application does not constitute a limitation on the scope of protection of the present application. Any various other corresponding changes and modifications made according to the technical concept of the present application should be included in the scope of protection of the claims of the present application.

Claims

1. A flow matching based image defogging enhancement and intelligent perception collaborative optimization method, characterized in that, The method comprises the steps of: obtaining a foggy input image, constructing an ordinary differential equation defining image transformation, the ordinary differential equation being defined by a fog-aware vector field; inputting the input image into the fog-aware vector field, the fog-aware vector field comprising: an atmospheric scattering purifier for extracting multi-scale features from the foggy input image to generate a clearer output, and a defogging-aware color lookup table for adaptively adjusting image colors through a nonlinear color transformation mechanism; solving the ordinary differential equation through an RK4 solver, and iteratively updating the foggy input image at each time step to ensure that the input image is smoothly and stably transformed from an initial foggy state to a clear image; A three-dimensional lookup table represents a three-dimensional lattice, composed of elements, denoted as where is the number of quantization levels for each color channel. Each element corresponding to the RGB color index and its transformed output RGB color The accuracy of this transformation is determined by The transformation of the defogging perceptual color lookup table is achieved through two key steps: lookup and trilinear interpolation; For an input RGB color , the coordinates in the lookup table are calculated as follows: ; where, is the maximum color value; the output RGB color is obtained by tri-linear interpolation from the 8 neighboring lattice vertices; finally, the defog-aware color lookup table procedure is represented as: ; wherein, is the result of color adjustment by the defogging perceptual color lookup table, is the input image, denotes a trilinear interpolation function, is the defogging perceptual color lookup table containing the sampled output RGB values.

2. The method of claim 1, wherein, the defogging process is controlled by the ordinary differential equation: ； wherein denotes a time step, is a time of an image, is a fog-aware vector field parameterized by is an initial time step, for the input image, the input image gradually transforms from a foggy state to a clear state along a probabilistic path by solving the ordinary differential equation.

3. The method of claim 2, wherein, the processing process of the fog-aware vector field on the input image is represented as: ; wherein, an output representing a haze-aware vector field, is a current input image, is an output image obtained by an atmospheric scattering purifier, is a result of color adjustment by a haze-aware color lookup table, is a scaling factor to control the intensity of color transformation.

4. The method of claim 3, wherein, an atmospheric scattering model is designed in the atmospheric scattering purifier and is represented as: ； wherein, is a clear image, is an input foggy image, is a learnable scaling parameter, is a learnable bias term with an initial value of 1.

5. The method of claim 4, wherein, by a convolutional neural network (CNN) for estimating directly from the input The defogged image can be obtained by the following way: ； wherein, is the current input image, is the output image obtained by the atmospheric scattering purifier, is the output of the enhanced convolutional neural network, which is composed of three down-sampling encoder modules, one spatial attention module and three up-sampling decoder modules: ； wherein, is the current input image, denotes a down-sampling encoder module, refers to an attention mechanism, denotes an up-sampling decoder module.

6. A flow matching based image defogging enhancement and intelligent perception collaborative optimization system, applied with the flow matching based image defogging enhancement and intelligent perception collaborative optimization method of claim 1, characterized in that, The method comprises the steps of: a data acquisition module for obtaining a foggy input image; an equation construction module for constructing an ordinary differential equation defining image transformation, the ordinary differential equation being defined by a fog-aware vector field; a fog-aware vector field comprising: an atmospheric scattering purifier for extracting multi-scale features from the foggy input image to generate a clearer output, and a defogging-aware color lookup table for adaptively adjusting image colors through a nonlinear color transformation mechanism; an image defogging module for solving the ordinary differential equation through an RK4 solver, and iteratively updating the foggy input image at each time step to ensure that the input image is smoothly and stably transformed from an initial foggy state to a clear image.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores computer executable instructions for causing a computer to execute the image defogging enhancement and intelligent perception collaborative optimization method based on flow matching according to any one of claims 1 to 5.

Images